OpenMV4 Plus utilises Edge Impulse to self-train neural networks for mask recognition.

Last April, our new product, the OpenMV4 H7 Plus, was launched. Today, I will introduce its new feature—using the EdgeImpulse online platform to independently train neural networks for classification and recognition.

In this tutorial, you will use machine learning to build a recognition system that can identify whether a person is wearing a mask through the OpenMV4 H7 Plus smart camera, achieving image classification.

EdgeImpulse is a website that provides online training services for neural network models on embedded devices. It is a partner of our OpenMV and both are official partners of STMicroelectronics. Currently, EdgeImpulse is free for our OpenMV users, allowing them to train neural network models online for OpenMV at no cost.

Adding neural networks to the embedded device OpenMV enables functions such as distinguishing poachers from elephants, performing quality control on factory production lines, and allowing remote-controlled model cars to drive autonomously.

In this tutorial, you will learn how to collect images to build a high-quality dataset, how to apply transfer learning to train a neural network, and how to deploy the system to OpenMV.

Additionally, we have provided video tutorials, which can be viewed by searching for "Training Neural Networks for Mask Recognition with OpenMV4 Plus" on the official website.

You can view the entire project content, including all code and models, at the following address: https://book.openmv.cc/project/mask.html

Using EdgeImpulse to train a neural network model online for OpenMV primarily involves the following four steps: dataset collection, uploading, training, and deployment.

01. Data Collection Dataset

In this tutorial, we will build a model capable of distinguishing whether a face is wearing a mask. You may also choose to classify other objects. To enable the operation of the machine learning model, you need to collect a large number of example images of faces both with and without masks. During training, these example images are used for the model's discrimination practice.

We need to utilise OpenMV IDE to collect our dataset. The steps for capturing images are as follows:

Create two categories:

Open the "Tools" menu in the OpenMV IDE, select "Dataset Editor," click "New Dataset," then create a new folder and open it. Inside the dataset, create two additional folders named "mask" (for saving photos of faces wearing masks) and "face" (for saving photos of faces without masks).

Collect images using OpenMV

First, we connect the OpenMV, click "Connect" in the IDE, then click "Run". The real-time image from the OpenMV can be seen in the Framebuffer. By clicking the photo button in the left menu bar, the OpenMV will automatically save this image, which becomes one data entry in the dataset.

First, we need to save the face of a person wearing a mask. Click the photo button, and the IDE will capture an image at the bottom, automatically naming it 00000. The next one will be 00001, and so forth.

Collect 200 photos of faces wearing masks (100 each of males and females) and 200 photos of faces not wearing masks (100 each of males and females) using OpenMV.

Note: Ensure photos are captured from various angles to guarantee the diversity of our training and learning.

If during the collection process you find that an image is not captured perfectly, you can right-click on the image and select delete.

02. Upload

After capturing images using OpenMV, you need to register an account and log in to the EdgeImpulse official website (https://edgeimpulse.com/) to begin uploading images. The steps for uploading images are as follows:

Create a new project on EdgeImpulse, click on "keys", select "API Key" and copy the "API Key". Use the API Key to establish connectivity between OpenMV IDE and Edge Impulse.

Select "Tools" – "Data Editor" – "Export" – "Upload" – "Upload via API Key" from the menu bar at the top of OpenMV IDE, then copy the "API Key" to upload.

During upload, the data will be automatically split into training and test sets. We recommend using the default ratio of "80% and 20%". We have collected over 400 facial images in total. Here, 80% of the Training Data means that 80% of your data is used for training by default, while the remaining 20% is allocated to the test set.

03. Training Dataset

Once the dataset is prepared, you can train it on the EdgeImpulse website interface.

Configuration Processing Module:

First, select "Impulse Design" to configure the processing module:

* Set the default image dimensions to "96 x 96".

* Select the "Images" module, indicating that we are conducting classification training on images (you can also use EdgeImpulse to classify sounds, videos, etc.);

* Select "Transfer Learning (Images)", set the learning model;

Select "Save impulse", and if "Successfully" is displayed, the configuration is successful.

Image Preprocessing:

Click "Image" in the left-hand menu, set the colour format to "RGB", then click "Save".

Next, select "Generate Features" to initiate the process. This will perform image preprocessing on over 400 data points, with a 3D visualisation of the complete dataset displayed on the right.

Configure the transfer learning model

Click "Transfer learning" in the left-hand menu, and select all default parameters for the settings, or adjust them as needed.

1. Set the default number of training cycles to 10

2. Set the learning rate to 0.0005.

3. You may choose to tick or leave unticked the "Data augmentation" option.

4. Set the default Minimum confidence rating to 0.8.

Click "Start training". The training process will take approximately four to five minutes.

After the model training is completed, you can view the accuracy, confusion matrix, and expected device performance.

* Test model:

Upon completion of training, we will utilise test data to evaluate the model.

Select "Model testing", tick the checkbox next to "Sample name", and click "Classify selected". The accuracy displayed here has reached 97%, which is quite remarkable for a model with such limited data.

The image highlighted in red indicates an uncertain classification. Click the three dots to the right of this image, select "Show classification", and you will enter the "Real-time Classification" interface, which contains more details about the file. This interface will help you determine the reason for the image's misclassification.

If the data falls outside all known clusters (wearing a mask/not wearing a mask), this may be data that does not match any previously observed classification—possibly due to hair covering a large portion of the face.

04. Running the Model on OpenMV

After completing the steps of training design, model training, and model validation, you can export this model to your OpenMV. Select "OpenMV," choose "Build" to generate, and it will automatically create three files.

* trained.tflite trained neural network model

* labels.txt two classification labels (face and mask)

* ei_image_classification.py is the code to be run on the OpenMV

Connect to OpenMV IDE and save these three files to the built-in Flash of OpenMV. Open the ei_image_classification.py file in OpenMV IDE to view the code we just generated in EdgeImpulse. Click run, and the operation results will be displayed in the serial terminal.

Alright, you have successfully added the capability for OpenMV to train neural networks on its own.

Looking forward to your results!